Beyond the AI hype: what should your business actually automate?
Almost nobody in a room of small business owners is behind on AI, and almost nobody is ahead. The Federal Reserve's 2025 Small Business Credit Survey found that 46% of small employer firms use AI in some way. Among those users, about half are still experimenting, 44% have built it partly into a business process, and 7% have built it in fully (Federal Reserve Banks, 2026 Report on Employer Firms). Do the arithmetic across all firms and roughly 20 in 100 have AI partly in their operation and about 3 in 100 have it fully in. Everyone can open these tools on their phone tonight. Almost nobody has gone past the chat window.
This post is the written version of a talk I am giving to the Savannah Area Chamber of Commerce today. If you were in the room, this is the leave-behind. If you were not, it is the whole argument in one read.
The vocabulary first
Six words cover everything: model, API, tool, agent, skill, harness. The model is the engine that answers. The API is the pipe to the vendor's servers. Tools let the model reach your calendar, inbox, and spreadsheets. An agent takes a goal instead of a prompt and runs a loop (plan, act, check, repeat) with a gate where a person approves anything that matters. A skill is the written playbook for one job. The harness is everything you would hand a new hire on day one, wrapped around the model: instructions, skills, tools, memory, permissions.
The full version with the drawing is in Six words that explain all of AI. The one-line test to carry around: if a vendor says agent and cannot show you the loop, the tools, and the gate, you are looking at a chatbot with a new label. Gartner estimated in June 2025 that only about 130 of the thousands of vendors claiming agentic capability actually have it (Gartner).
Three decisions, like the legs of a stool
Every business that uses AI makes three decisions. Most make the first two by accident, by signing up for whatever came first. The ones getting value made them on purpose. I think of them as the legs of a stool because a stool with two legs falls over.
Leg one: where your AI runs
Rent the model through a vendor's API, or run it on hardware you own. Renting is cheap to start, fast, and maintained by someone else; your data leaves the building and the cost recurs. Owning keeps the data in the office and drops the monthly bill to electricity; it costs money up front and someone has to maintain the machine. Open-weight models you can run locally trail the best cloud models by about four months on average (Epoch AI, May 2026), which is plenty for repetitive work and not enough for the hardest reasoning.
The decision turns on five things: how sensitive your data is, how much capital you have now, how much you already spend on AI each month, how many systems the AI needs to touch, and who will maintain it. For most small businesses today the honest answer is rent, on a business-tier account, for everything except the most sensitive work. For some, the accountants and medical practices and firms already spending real money, the honest answer is a machine of their own for that work. The full decision matrix and current costs are in Where should your AI run?, and the local-machine path is what Maai Machines exists to build.
Leg two: what data it is allowed to touch
Sort your data into three tiers. Green is public: any tool, any account. Yellow is confidential (pricing, processes, client emails): business-tier accounts only, where the vendor agrees in writing not to train on your data, or your own hardware. Red is regulated (Social Security numbers, health records, payroll, financials): a machine you control, or it does not touch AI.
The part most owners do not know is that the $20 personal plan and the business plan from the same vendor treat your data differently. Anthropic, OpenAI, and Google all follow the same shape: on consumer plans, training is on by default and you have to turn it off; on business plans and the API, training is off by contract (Anthropic; OpenAI). Two moves this week: put your team on business accounts, and keep red data off the cloud. Then experiment freely with everything else.
The second data rule is that structure going in decides quality coming out. The most valuable data most businesses own is conversational (email, notes, transcripts) and almost none of it is organized. An AI handed a messy pile gives a messy answer. Gartner predicted in February 2025 that through 2026 organizations will abandon 60% of AI projects not supported by AI-ready data (Gartner). The model is rarely the problem. The pile is. Both rules are worked through in The three-tier data rule.
Leg three: how you break the work into pieces
This leg is a way of thinking, and the last time this happened is the best way to explain it. Mainframes were for large companies with volume to justify them. The personal computer put the same capability on every desk, and for years most of it ran a word processor. The value arrived when software digitized specific work: the spreadsheet, then accounting, then scheduling and the CRM. Hardware made it possible. Software made it pay. The businesses that got the most out of it were the ones that rethought how the work flowed, instead of doing the old process on a new screen.
AI multiplies both layers at once. The speed is real. The complexity did not go away. The same discipline that decided who won with the PC decides who wins now.
Decompose the work. Take one process your business repeats every week and write every step down: who does it, how long it takes, what goes in, what comes out, what data it touches. Operations people call it a brown-paper exercise because you tape paper to a wall and write on it. It takes an hour, and almost nobody has done it for their own business. Then tag each step one of three ways: needs a human, AI can do it with a human checking, AI can do it alone. Most steps in most processes land in the middle.
Ask two questions of each step. Does AI do this well when you test it on ten real examples, not when the vendor says so? And what does it cost if it gets one wrong? Does well and cheap if wrong: automate it and spot-check on Fridays. Does well but costly if wrong: the agent drafts, a person approves. Not there yet and cheap: test small, one person, one month. Not there yet and costly: keep it human. For your first one, pick a step that costs three or more hours a week and that you can check in under two minutes. Three hours makes it worth your attention. Two minutes to check means you will check it.
Climb the ladder. A moment: one task, one prompt, you check it. A flow: several steps chained, you still press go. A skill: the flow written down so it runs the same way every time. An agent: it owns the job, runs on a schedule, asks before anything goes out. A team: several agents, one lead, working from a brief. Almost everyone is on rung one. The output comes back on rungs two and three.
Measure it, because feelings lie
The evidence on whether AI pays is mixed, and the mix is instructive. METR ran a randomized trial with 16 experienced open-source developers working on code they knew well. With AI tools they took 19% longer to finish tasks, while believing afterward that they had been 20% faster (METR, July 2025). A study of 5,179 customer support agents found output rose 14% on average when an AI assistant was built into the workflow, and 34% for newer staff (Brynjolfsson, Li, and Raymond, NBER). A Danish study of 25,000 workers in 11 occupations using chatbots mostly on their own found average time savings of 2.8% and no measurable effect on earnings or hours (Humlum and Vestergaard, 2025).
Read together: built into the workflow and measured, it pays. Dabbled with, it does not. And people cannot tell the difference by feel. So measure three things before and after: how long the process takes, how often it is wrong, and how many hours moved to work only a person can do.
Then think past this quarter. A business is valued on a multiple of its earnings, and a point of margin you add by removing repeated manual work shows up in that valuation every year after. If you are thinking about selling, bringing in a successor, or borrowing against the business in the next several years, lenders and buyers price in documented, repeatable processes. That is the logic of the thing, not a statistic, and it is the reason putting this off has a cost even when the monthly savings look small.
Software stays
You have heard you can build your own CRM in a weekend and cancel the subscription. You can build something in a weekend. You should not. The platforms you pay for have a decade of features, security, and edge cases built in. What changed is that building on top of them got much easier. Add the feature. Do not rebuild the house. Businesses that tear out their software spend two months breaking things and two months not serving customers.
Five starts, one a day
The Fed asked firms with no plans to use AI why not; over half said it does not apply to their business (Federal Reserve Banks). In my experience the barrier is rarely money. It is knowing where to start. None of these cost a dollar.
- Monday: fix the accounts. Move anyone using AI for work onto a business-tier account and turn off training on any personal account that remains.
- Tuesday: sort the data. Hand out the three-tier card. Everyone tags what they touch: green, yellow, red.
- Wednesday: decompose one process. One hour. Every step, who does it, how long, what data it touches.
- Thursday: pick a first hire. One step, three or more hours a week, checkable in two minutes. Try it with a personal AI assistant this week. Chain it next month.
- Friday: email your insurance broker. Ask whether your policy excludes losses from AI. It is a yes or no, and you want it before you need it.
What happens next
The ideas here are five years old. The parts just arrived. On Friday, November 6, I am hosting this month's coffee hours: one-on-one or small group, an hour, a coffee, no pitch. Bring the process from your worksheet and leave with a plan for it. Sign up here. The deck, the six-words card, the three-tier card, and the worksheet go to new newsletter subscribers, with one short note a week after that.
Go build something.
Sources: Federal Reserve Banks, 2026 Report on Employer Firms; Gartner, June 25, 2025; Gartner, February 26, 2025; Epoch AI, May 29, 2026; Anthropic consumer terms update; OpenAI business data; METR, July 10, 2025; Brynjolfsson, Li, Raymond, NBER WP 31161; Humlum and Vestergaard, BFI WP 2025-56. Figures checked October 7, 2026.